RF Gauntlet

A hostile-but-legal RF environment for connectivity testing.

Field connectivity failures almost never come from a single broken device. They come from density: thirty APs in an apartment building, a parking lot full of cars running phone-as-key, a conference hall of BLE beacons. Every interferer in this rig is 100% within its radio spec. Only the combination — and the duty cycle — is worst-case.

The goal is two-sided:

1. Survivability — put our device (wearable + paired phone) in the middle of that and prove the link stays up.
2. Citizenship — prove our device isn't contributing to the hell. Passive observers capture everything the DUT transmits and score it against docs/dut-citizenship.md. A device that holds its link by shouting over everyone else fails.

Principle

• In-spec only. No jammers, no out-of-band tricks, no protocol violations. If a device can't legally do it, it doesn't go in the chamber.
• Worst-case duty cycle. Every node runs at the most aggressive behavior its spec allows: minimum advertising intervals, back-to-back inquiry trains, 100%-duty-cycle scanning, beacons at the lowest basic rate.
• Real things. Each interferer models something that actually exists in the field: the car doing phone-as-key in the garage, the AP farm in the apartment block, the Auracast broadcast at the airport.
• LLM-conducted. A conductor model composes scenarios from the interferer catalog, executes them, watches DUT link metrics, escalates hostility until something breaks, and writes up the breaking point with evidence.

Architecture

                        +------------------+
                        |  LLM conductor   |  plans, executes, scores
                        +--------+---------+
                                 | control plane (HTTP/JSON-RPC)
        +------------------------+------------------------+
        |              RF chamber / quiet room             |
        |                                                  |
        |  [BR/EDR nodes]  [BLE nodes]  [LE Audio] [Wi-Fi] |
        |  inquiry looper   adv spammers  Auracast  AP farm |
        |  page stormer     active scan   (BIS)     slow    |
        |  A2DP hog        conn saturator            client |
        |                  phone-key sim  [15.4]    probe   |
        |                  ext-adv spam    zigbee    stormer |
        |                                                  |
        |   DUT: wearable + paired phone                   |
        |   (behind programmable attenuator)               |
        +--------------------------------------------------+

The conductor never touches firmware. It drives node agents through the control plane: start/stop nodes, set parameters, set chamber path loss, sample DUT metrics, and render a verdict.

Repo map

• docs/bom.md — parts and device list
• docs/interferers.md — interferer catalog: spec basis, worst-case params, hardware, expected victim impact
• docs/scenario-dsl.md — scenario YAML schema
• docs/llm-driver.md — the conductor spec: role, tool interface, control loop, scoring, safety rails
• scenarios/ — example scenarios (smoke.yaml, worst-case.yaml)
• driver/ — control-plane agent contract + orchestrator skeleton

Quickstart

1. Build the fleet per docs/bom.md and flash node firmware.
2. Place nodes in the chamber; DUT behind the programmable attenuator.
3. Bring up the control plane: one HTTP agent per node + driver/orchestrator.py.
4. Run the smoke scenario, then let the conductor escalate:

python driver/orchestrator.py --scenario scenarios/smoke.yaml

Safety

• Prefer a shielded chamber or shielded room. Open-air runs stay within ISM band rules; node TX power caps are enforced by the agent contract, not by asking nicely.
• The conductor MUST NOT exceed per-node TX caps or run longer than the scenario's time box. See docs/llm-driver.md safety rails.
• DUT protection: abort criteria (excessive disconnect churn) exist so a bad scenario can't wedge the DUT in a reconnect loop indefinitely.